Optimal Fusion Weights Form Intervals, Not Points: A Latency-Aware Study of Query-Adaptive WRRF

Abstract

We build ten query-adaptive weighted reciprocal rank fusion (WRRF) methods for production search, showing they outperform standard RRF. We find that optimal fusion weights form intervals rather than single points, helping explain why RRF remains near-optimal for 10–38% of queries. We introduce an oracle/headroom-recovery framework with latency-aware tiering to make these gains practical in production, and evaluate across four sparse and dense retriever configurations (BM25/RM3, MiniLM, Qwen3).

Iris Ma

Software Engineering Ph.D. Candidate

My research focuses on AI for reliable software engineering, including LLM memorization, hallucination detection, retrieval-augmented generation, and verified program synthesis.